Multiprocessing in MAPIE fails with Python 3.12/3.13

[Valentin-Laurent]

Multiprocessing (ie: n_jobs=-1) in MAPIE fails when using python 3.13 and 3.12 (maybe it was failing silently for older versions).

This is the case at least when using _MapieClassifier, and also most probably using _MapieRegressor, TimeSeriesRegressor, and PrecisionRecallController.

Minimal code example:

from sklearn.datasets import make_classification
from mapie.classification import _MapieClassifier

n_classes = 4
X, y = make_classification(
    n_samples=500,
    n_features=10,
    n_informative=3,
    n_classes=n_classes,
)

mapie_clf_multi = _MapieClassifier(n_jobs=-1)
mapie_clf_multi.fit(X, y)
mapie_clf_multi.predict(X)

Returns ChildProcessError: [Errno 10] No child processes [..] (in the ResourceTracker module)

When running the test suite, the test_results_single_and_multi_jobs tests raise exceptions. Those exceptions are displayed but ignored, so the test suite passes, but the code would fail in a regular environnement.

[ramirorcr]

MapieClassifier is now a private class and no longer exposed to the user. I tried to reproduce the error with the new alternative CrossConformalClassifier, but no errors were raised. Can we reconsider lifting the restriction on the Python version?

[Valentin-Laurent]

Hello @ramirorcr. Thank you for spotting this. Indeed I can't reproduce the error using my code snippet (edited to take into account recent API changes).

This issue has a high priority in our backlog anyways, so I'm currently exploring why I got this error when I reported the problem. I'll let you know my conclusions.

In the meantime, if you want you can use pip install mapie==1.0.1 --ignore-requires-python to install MAPIE in a Python 3.12+ environnement.

[Valentin-Laurent]

Hello @ramirorcr.

We didn't manage to reproduce the issue. So support for Python 3.12 and above will be live in the next release.